Has anybody worked with AI in some capacity?

On or with? If it's with, then yes, although not very much. As a software developer, I find it occasionally useful for generating helper scripts, although it has a tendency to generate seemingly legit but actually broken nonsense for anything non-trivial. It's much more useful for searching things that have traditionally been hard to search, especially with the terrible quality of search engines these days.

As for working on AI in some capacity, the answer is no for the purposes of this thread (I have meddled a bit in game AI, but even that's to some very basic capacity many, many years ago now). I do have a basic understanding of some of the core principles, but naturally my grasp on them is fairly weak because I haven't actually got my hands dirty. I think we have a member here that is working on AI, if I recall correctly?

Not the neural network kind where what influence a decision is sort of hidden. Have built, or rather improved on, a few text-editing tools for summarization and ease of read. But they are only AI in the sense that they try to mimic intelligence, but was built around rules instead of statistics. Built an NPC unit in a game like asteroid. Back then they where often referred to as "AI".

Studied AI a bit back at university but since the tech improves so fast not much of that would even qualify for the AI label. Built algorithms for "smart" vacuum cleaning, set up information hierarchies and what the different building blocks of information (or data) are. Also built and trained a few simple neural networks for pattern recognition, weighting nodes, training on simple data and executing tasks like if there are four black dots in a row the next dot will be white.

Not in the last decade. Before that I worked on GPS navigation. Which had what was then called AI for route finding, dead reckoning, voice recognition, address matching, basically a lot of fuzzy matching. As well as tracking cell phone tower data to predict / determine the locations of traffic jams and feed back the actual traffic speed on major roads.

Of course of all those the rule based 'AI' was most reliable, the neural net parts (voice recognition) the least, or rather the hardest to fix/improve. With limited processing power it all had to be either rule based or simple neural nets. Server based is still not great though, Siri and my TV often still don't understand my Dutch accent. :/

Anyway none of those were in any way threatening, but we did had some moral concerns about tracking user data and the whole scale tracking of all cell phone location data.

Never used AI for code generation, but if AI could help find bugs, that could be useful. Not just make the program crash but actually find why it crashes under what circumstances. That's the hardest part of the job, those infernal 'can't replicate' easily crashes and unintended behavior. Which still happens when everything is rule based. Murphy's law in software "Everything that can go wrong, eventually will go wrong". We proved that rule all the time lol.

One of the worst was tracking down unexpected slow down and increase in memory use in the routing engine. Eventually it turned out to be caused by a unique situation in the data network. A logging area in Germany somehow had an exact geometric pattern of exactly equal length roads, a grid pattern. A condition in the code kept both options open when search paths reach the next crossing with the exact same value. Since it was a large grid pattern like a checkerboard, it basically replicated the checkerboard problem. Doubling the open options (paths to investigate) at every intersection. Yet only when this area fell within search range it would start exploring there and thus slowing down the useful search paths the longer the search had to go on from there. Either completing it slowly or eventually running out of memory.

Another one was just if not harder to find. A condition where left turn prohibitions from all sides on a crossing could create another looping condition. Both bugs and others were eventually found with visualization of the search tree. Then spotting by eye where suspicious behavior occurs, activity that goes on too long in an area, areas that aren't reached, unexpected jumps etc. AI could be useful to spot things like that.

Visualization of code running has been very useful for optimizing disk access as well. Finding patterns in data access to organize data more efficiently, reducing the number of reads, block sizes to use, optimizing what should stay in memory and for how long. Using the human mind for pattern recognition. AI used to improve compression and data organization would be useful. Of course at one point Huffman coding was considered AI.


Anyway we were all about reducing costs and optimization, this just doesn't compute to me:

https://www.eurogamer.net/google-reaches-grim-ai-milestone-replicates-playable-version-of-doom-entirely-by-algorithm

How to make things less efficient...

Those diffusion models are the opposite of what you're suggesting.

The RL plays DOOM and the neural model learns to match the controls with the images. The neural models uses a diffusion approach, like Stable Diffusion image generator and learnt from 900-millions frames that the RL generated in its play-throughs.

First you got to build it, then play it, then the diffusion model can make a worse running copy...

Then copy enough games and you have an FPS generator. 

No, an fps replicator ;)

AI can be useful to optimize games. Brute force guided search what to nip and tuck to maintain stable frame rates. The more games it optimizes the better it could get at it. I guess AI upscaling is already a form of that.

If Ubisoft trained an Ai on all of there IP to spit out some form of a game on the cheap, would it be any different from a team making a new game. 

booyah

how ironic, as judgement day was just a few days ago.... soon the bots and xbots and pspbots are gonna come for us all